inline
void print_bitmap( std::ostream & s , const hwloc_const_bitmap_t bitmap )
{
s << "{" ;
for ( int i = hwloc_bitmap_first( bitmap ) ;
-1 != i ; i = hwloc_bitmap_next( bitmap , i ) ) {
s << " " << i ;
}
s << " }" ;
}
/******************* FUNCTION *********************/
int TopoHwloc::getFirstBitInBitmap(hwloc_bitmap_t bitmap) const
{
int last = hwloc_bitmap_last(bitmap);
int current = hwloc_bitmap_first(bitmap);
assert(current != -1);
while (current != last)
{
if (hwloc_bitmap_isset(bitmap,current))
break;
current = hwloc_bitmap_next(bitmap,current);
}
return current;
}
/* NTH: this is no longer used but may be used if we can determine the binding policy*/
static int mca_sbgp_map_to_logical_socket_id(int *socket)
{
int ret = OMPI_SUCCESS;
hwloc_obj_t obj;
hwloc_obj_t first_pu_object;
hwloc_bitmap_t good;
int pu_os_index = -1, my_logical_socket_id = -1;
int this_pus_logical_socket_id = -1;
*socket = my_logical_socket_id;
/* bozo check */
if (NULL == opal_hwloc_topology) {
return OPAL_ERR_NOT_INITIALIZED;
}
good = hwloc_bitmap_alloc();
if (NULL == good) {
return OPAL_ERR_OUT_OF_RESOURCE;
}
/* get this process' CPU binding */
if( 0 != hwloc_get_cpubind(opal_hwloc_topology,good, 0)){
/* report some error */
BASESMSOCKET_VERBOSE(10, "The global variable opal_hwloc_topology appears not to have been initialized\n");
hwloc_bitmap_free(good);
return OMPI_ERROR;
}
/* find the first logical PU object in the hwloc tree */
first_pu_object = hwloc_get_obj_by_type(opal_hwloc_topology, HWLOC_OBJ_PU, 0);
/* get the next bit in the bitmap (note: if pu_os_index == -1, then the
* first bit is returned
*/
/* traverse the hwloc tree */
while( -1 != (pu_os_index = hwloc_bitmap_next(good, pu_os_index) ) ) {
/* Traverse all PUs in the machine in logical order, in the simple case
* there should only be a single PU that this process is bound to, right?
*
*/
for( obj = first_pu_object; obj != NULL; obj = obj->next_cousin ) {/* WTF is a "next_cousin" ? */
/* is this PU the same as the bit I pulled off the mask? */
if( obj->os_index == (unsigned int) pu_os_index) {
/* Then I found it, break out of for loop */
break;
}
}
if( NULL != obj) {
/* if we found the PU, then go upward in the tree
* looking for the enclosing socket
*/
while( (NULL != obj) && ( HWLOC_OBJ_SOCKET != obj->type) ){
obj = obj->parent;
}
if( NULL == obj ) {
/* then we couldn't find an enclosing socket, report this */
} else {
/* We found the enclosing socket */
if( -1 == my_logical_socket_id ){
/* this is the first PU that I'm bound to */
this_pus_logical_socket_id = obj->logical_index;
my_logical_socket_id = this_pus_logical_socket_id;
} else {
/* this is not the first PU that I'm bound to.
* Seems I'm bound to more than a single PU. Question
* is, am I bound to the same socket??
*/
/* in order to get rid of the compiler warning, I had to cast
* "this_pus_logical_socket_id", at a glance this seems ok,
* but if subgrouping problems arise, maybe look here. I shall
* tag this line with the "mark of the beast" for grepability
* 666
*/
if( (unsigned int) this_pus_logical_socket_id != obj->logical_index ){
/* 666 */
/* Then we're bound to more than one socket...fail */
this_pus_logical_socket_id = -1;
my_logical_socket_id = -1;
break;
}
}
}
}
/* end while */
}
*socket = my_logical_socket_id;
hwloc_bitmap_free(good);
return ret;
}
int main(void)
{
hwloc_bitmap_t set;
/* check an empty bitmap */
set = hwloc_bitmap_alloc();
assert(hwloc_bitmap_to_ulong(set) == 0UL);
assert(hwloc_bitmap_to_ith_ulong(set, 0) == 0UL);
assert(hwloc_bitmap_to_ith_ulong(set, 1) == 0UL);
assert(hwloc_bitmap_to_ith_ulong(set, 23) == 0UL);
/* check a non-empty bitmap */
hwloc_bitmap_from_ith_ulong(set, 4, 0xff);
assert(hwloc_bitmap_to_ith_ulong(set, 4) == 0xff);
assert(hwloc_bitmap_to_ulong(set) == 0UL);
assert(hwloc_bitmap_to_ith_ulong(set, 0) == 0UL);
assert(hwloc_bitmap_to_ith_ulong(set, 1) == 0UL);
assert(hwloc_bitmap_to_ith_ulong(set, 23) == 0UL);
/* check a zeroed bitmap */
hwloc_bitmap_zero(set);
assert(hwloc_bitmap_to_ulong(set) == 0UL);
assert(hwloc_bitmap_to_ith_ulong(set, 0) == 0UL);
assert(hwloc_bitmap_to_ith_ulong(set, 1) == 0UL);
assert(hwloc_bitmap_to_ith_ulong(set, 4) == 0UL);
assert(hwloc_bitmap_to_ith_ulong(set, 23) == 0UL);
hwloc_bitmap_free(set);
/* check a full bitmap */
set = hwloc_bitmap_alloc_full();
assert(hwloc_bitmap_to_ulong(set) == ~0UL);
assert(hwloc_bitmap_to_ith_ulong(set, 0) == ~0UL);
assert(hwloc_bitmap_to_ith_ulong(set, 1) == ~0UL);
assert(hwloc_bitmap_to_ith_ulong(set, 23) == ~0UL);
/* check a almost full bitmap */
hwloc_bitmap_set_ith_ulong(set, 4, 0xff);
assert(hwloc_bitmap_to_ith_ulong(set, 4) == 0xff);
assert(hwloc_bitmap_to_ulong(set) == ~0UL);
assert(hwloc_bitmap_to_ith_ulong(set, 0) == ~0UL);
assert(hwloc_bitmap_to_ith_ulong(set, 1) == ~0UL);
assert(hwloc_bitmap_to_ith_ulong(set, 23) == ~0UL);
/* check a almost empty bitmap */
hwloc_bitmap_from_ith_ulong(set, 4, 0xff);
assert(hwloc_bitmap_to_ith_ulong(set, 4) == 0xff);
assert(hwloc_bitmap_to_ulong(set) == 0UL);
assert(hwloc_bitmap_to_ith_ulong(set, 0) == 0UL);
assert(hwloc_bitmap_to_ith_ulong(set, 1) == 0UL);
assert(hwloc_bitmap_to_ith_ulong(set, 23) == 0UL);
hwloc_bitmap_free(set);
/* check ranges */
set = hwloc_bitmap_alloc();
assert(hwloc_bitmap_weight(set) == 0);
/* 23-45 */
hwloc_bitmap_set_range(set, 23, 45);
assert(hwloc_bitmap_weight(set) == 23);
/* 23-45,78- */
hwloc_bitmap_set_range(set, 78, -1);
assert(hwloc_bitmap_weight(set) == -1);
/* 23- */
hwloc_bitmap_set_range(set, 44, 79);
assert(hwloc_bitmap_weight(set) == -1);
assert(hwloc_bitmap_first(set) == 23);
assert(!hwloc_bitmap_isfull(set));
/* 0- */
hwloc_bitmap_set_range(set, 0, 22);
assert(hwloc_bitmap_weight(set) == -1);
assert(hwloc_bitmap_isfull(set));
/* 0-34,57- */
hwloc_bitmap_clr_range(set, 35, 56);
assert(hwloc_bitmap_weight(set) == -1);
assert(!hwloc_bitmap_isfull(set));
/* 0-34,57 */
hwloc_bitmap_clr_range(set, 58, -1);
assert(hwloc_bitmap_weight(set) == 36);
assert(hwloc_bitmap_last(set) == 57);
assert(hwloc_bitmap_next(set, 34) == 57);
/* 0-34 */
hwloc_bitmap_clr(set, 57);
assert(hwloc_bitmap_weight(set) == 35);
assert(hwloc_bitmap_last(set) == 34);
/* empty */
hwloc_bitmap_clr_range(set, 0, 34);
assert(hwloc_bitmap_weight(set) == 0);
assert(hwloc_bitmap_first(set) == -1);
hwloc_bitmap_free(set);
return 0;
}
int main(void)
{
hwloc_bitmap_t set;
int i, cpu, expected_cpu = 0;
/* empty set */
set = hwloc_bitmap_alloc();
assert(hwloc_bitmap_first(set) == -1);
assert(hwloc_bitmap_last(set) == -1);
assert(hwloc_bitmap_next(set, 0) == -1);
assert(hwloc_bitmap_next(set, -1) == -1);
assert(hwloc_bitmap_weight(set) == 0);
/* full set */
hwloc_bitmap_fill(set);
assert(hwloc_bitmap_first(set) == 0);
assert(hwloc_bitmap_last(set) == -1);
assert(hwloc_bitmap_next(set, -1) == 0);
assert(hwloc_bitmap_next(set, 0) == 1);
assert(hwloc_bitmap_next(set, 1) == 2);
assert(hwloc_bitmap_next(set, 2) == 3);
assert(hwloc_bitmap_next(set, 30) == 31);
assert(hwloc_bitmap_next(set, 31) == 32);
assert(hwloc_bitmap_next(set, 32) == 33);
assert(hwloc_bitmap_next(set, 62) == 63);
assert(hwloc_bitmap_next(set, 63) == 64);
assert(hwloc_bitmap_next(set, 64) == 65);
assert(hwloc_bitmap_next(set, 12345) == 12346);
assert(hwloc_bitmap_weight(set) == -1);
/* custom sets */
hwloc_bitmap_zero(set);
hwloc_bitmap_set_range(set, 36, 59);
assert(hwloc_bitmap_first(set) == 36);
assert(hwloc_bitmap_last(set) == 59);
assert(hwloc_bitmap_next(set, -1) == 36);
assert(hwloc_bitmap_next(set, 0) == 36);
assert(hwloc_bitmap_next(set, 36) == 37);
assert(hwloc_bitmap_next(set, 59) == -1);
assert(hwloc_bitmap_weight(set) == 24);
hwloc_bitmap_set_range(set, 136, 259);
assert(hwloc_bitmap_first(set) == 36);
assert(hwloc_bitmap_last(set) == 259);
assert(hwloc_bitmap_next(set, 59) == 136);
assert(hwloc_bitmap_next(set, 259) == -1);
assert(hwloc_bitmap_weight(set) == 148);
hwloc_bitmap_clr(set, 199);
assert(hwloc_bitmap_first(set) == 36);
assert(hwloc_bitmap_last(set) == 259);
assert(hwloc_bitmap_next(set, 198) == 200);
assert(hwloc_bitmap_next(set, 199) == 200);
assert(hwloc_bitmap_weight(set) == 147);
i = 0;
hwloc_bitmap_foreach_begin(cpu, set) {
if (0 <= i && i < 24)
expected_cpu = i + 36;
else if (24 <= i && i < 87)
expected_cpu = i + 112;
else if (87 <= i && i < 147)
expected_cpu = i + 113;
assert(expected_cpu == cpu);
i++;
} hwloc_bitmap_foreach_end();
hwloc_bitmap_free(set);
return 0;
}
void computeCPUOMP(int threadId, expression_type * expr, im_type * im, element_iterator * elt_it, std::vector<std::pair<element_iterator, element_iterator> > * elts)
{
char * a;
int cid;
std::ostringstream oss;
#if 0
hwloc_cpuset_t set = nullptr;
/* get a cpuset object */
set = hwloc_bitmap_alloc();
/* Get the cpu thread affinity info of the current process/thread */
hwloc_get_cpubind(Environment::getHwlocTopology(), set, 0);
hwloc_bitmap_asprintf(&a, set);
oss << a;
free(a);
cid = hwloc_bitmap_first(set);
oss << "(";
while(cid != -1)
{
oss << cid << " ";
cid = hwloc_bitmap_next(set, cid);
}
oss << ")|";
std::cout << Environment::worldComm().rank() << "|" << M_threadId << " " << oss.str() << std::endl;
/* Get the latest core location of the current process/thread */
hwloc_get_last_cpu_location(Environment::getHwlocTopology(), set, 0);
hwloc_bitmap_asprintf(&a, set);
oss << a;
free(a);
cid = hwloc_bitmap_first(set);
oss << "(";
while(cid != -1)
{
oss << cid << " ";
cid = hwloc_bitmap_next(set, cid);
}
oss << ");";
std::cout << Environment::worldComm().rank() << "|" << M_threadId << " " << oss.str() << std::endl;
#endif
#if defined(FEELPP_HAS_HARTS)
perf_mng.init("cpu") ;
perf_mng.start("cpu") ;
perf_mng.init("1.1") ;
perf_mng.init("1.2") ;
perf_mng.init("2.1") ;
perf_mng.init("2.2") ;
perf_mng.init("3") ;
#endif
//M_gm((*elt_it)->gm());
gm_ptrtype gm = (*elt_it)->gm();
//M_geopc(new typename eval::gmpc_type( M_gm, im->points() ));
typename eval::gmpc_ptrtype __geopc( new typename eval::gmpc_type(gm, im->points()) );
//M_c(new gmc_type( M_gm, *(*elt_it), M_geopc ));
gmc_ptrtype __c( new gmc_type( gm, *(*elt_it), __geopc ) );
//M_expr( (*expr), map_gmc_type( fusion::make_pair<vf::detail::gmc<0> >( M_c ) ) );
eval_expr_type __expr( (*expr), map_gmc_type( fusion::make_pair<vf::detail::gmc<0> >( __c ) ) );
for (int i = 0; i < elts->size(); i++)
{
/*
std::cout << Environment::worldComm().rank() << " nbItems: " << elts->size()
<< " nbElts " << std::distance(elts->at(i), elts->at(i+1))
<< " 1st id " << elts->at(i)->id() << std::endl;
*/
//std::cout << Environment::worldComm().rank() << "|" << theadId << " fid=" elts.at(i).first.id() << std::endl;
for ( auto _elt = elts->at(i).first; _elt != elts->at(i).second; ++_elt )
{
//perf_mng.start("1.1") ;
__c->update( *_elt );
//perf_mng.stop("1.1") ;
//perf_mng.start("1.2") ;
map_gmc_type mapgmc( fusion::make_pair<vf::detail::gmc<0> >( __c ) );
//perf_mng.stop("1.2") ;
//perf_mng.start("2.1") ;
__expr.update( mapgmc );
//perf_mng.stop("2.1") ;
//perf_mng.start("2.2") ;
im->update( *__c );
//perf_mng.stop("2.2") ;
//perf_mng.start("3") ;
for ( uint16_type c1 = 0; c1 < eval::shape::M; ++c1 )
{
for ( uint16_type c2 = 0; c2 < eval::shape::N; ++c2 )
{
M_ret( c1,c2 ) += (*im)( __expr, c1, c2 );
}
}
//perf_mng.stop("3") ;
}
}
#if defined(FEELPP_HAS_HARTS)
perf_mng.stop("cpu") ;
M_cpuTime = perf_mng.getValueInSeconds("cpu");
#endif
}
void computeCPU(DataArgsType& args)
{
char * a;
int cid;
hwloc_cpuset_t set = nullptr;
std::ostringstream oss;
/* This initialization takes some time */
/* When using hartsi, the object instanciation is done when creating tasks */
/* and this is not a parallel section, thus we lose time in initialization */
/* doing it the computation step allows to incorporate this init time in the parallel section */
/*
M_threadId( threadId ),
M_gm( new gm_type( *_elt.gm() ) ),
M_geopc( new gmpc_type( M_gm, _im.points() ) ),
M_c( new gmc_type( M_gm, _elt, M_geopc ) ),
M_expr( _expr, map_gmc_type( fusion::make_pair<vf::detail::gmc<0> >( M_c ) ) ),
M_im( _im ),
M_ret( eval::matrix_type::Zero() ),
M_cpuTime( 0.0 )
*/
#if 0
/* get a cpuset object */
set = hwloc_bitmap_alloc();
/* Get the cpu thread affinity info of the current process/thread */
hwloc_get_cpubind(Environment::getHwlocTopology(), set, 0);
hwloc_bitmap_asprintf(&a, set);
oss << a;
free(a);
cid = hwloc_bitmap_first(set);
oss << "(";
while(cid != -1)
{
oss << cid << " ";
cid = hwloc_bitmap_next(set, cid);
}
oss << ")|";
std::cout << Environment::worldComm().rank() << "|" << M_threadId << " " << oss.str() << std::endl;
/* Get the latest core location of the current process/thread */
hwloc_get_last_cpu_location(Environment::getHwlocTopology(), set, 0);
hwloc_bitmap_asprintf(&a, set);
oss << a;
free(a);
cid = hwloc_bitmap_first(set);
oss << "(";
while(cid != -1)
{
oss << cid << " ";
cid = hwloc_bitmap_next(set, cid);
}
oss << ");";
std::cout << Environment::worldComm().rank() << "|" << M_threadId << " " << oss.str() << std::endl;
#endif
perf_mng.init("1.1") ;
perf_mng.init("1.1") ;
perf_mng.init("2.1") ;
perf_mng.init("2.2") ;
perf_mng.init("3") ;
/* free memory */
if(set != nullptr)
{
hwloc_bitmap_free(set);
}
//perf_mng.init("data") ;
//perf_mng.start("data") ;
// DEFINE the range to be iterated on
std::vector<std::pair<element_iterator, element_iterator> > * elts =
args.get("elements")->get<std::vector<std::pair<element_iterator, element_iterator> > >();
int * threadId = args.get("threadId")->get<int>();
expression_type * expr = args.get("expr")->get<expression_type>();
im_type * im = args.get("im")->get<im_type>();
element_iterator * elt_it = args.get("elt")->get<element_iterator>();
//M_gm((*elt_it)->gm());
gm_ptrtype gm = (*elt_it)->gm();
//M_geopc(new typename eval::gmpc_type( M_gm, im->points() ));
typename eval::gmpc_ptrtype __geopc( new typename eval::gmpc_type(gm, im->points()) );
//M_c(new gmc_type( M_gm, *(*elt_it), M_geopc ));
gmc_ptrtype __c( new gmc_type( gm, *(*elt_it), __geopc ) );
//M_expr( (*expr), map_gmc_type( fusion::make_pair<vf::detail::gmc<0> >( M_c ) ) );
eval_expr_type __expr( (*expr), map_gmc_type( fusion::make_pair<vf::detail::gmc<0> >( __c ) ) );
//perf_mng.stop("data");
perf_mng.init("cpu") ;
perf_mng.start("cpu") ;
for (int i = 0; i < elts->size(); i++)
{
//std::cout << Environment::worldComm().rank() << " nbItems: " << elts->size() << " nbElts " << std::distance(elts->at(i), elts->at(i+1)) << std::endl;
for ( auto _elt = elts->at(i).first; _elt != elts->at(i).second; ++_elt )
{
//perf_mng.start("1.1") ;
//M_c->update( *_elt );
__c->update( *_elt );
//perf_mng.stop("1.1") ;
//perf_mng.start("1.2") ;
map_gmc_type mapgmc( fusion::make_pair<vf::detail::gmc<0> >( __c ) );
//perf_mng.stop("1.2") ;
//perf_mng.start("2.1") ;
__expr.update( mapgmc );
//perf_mng.stop("2.1") ;
//perf_mng.start("2.2") ;
im->update( *__c );
//perf_mng.stop("2.2") ;
//perf_mng.start("3") ;
for ( uint16_type c1 = 0; c1 < eval::shape::M; ++c1 )
{
for ( uint16_type c2 = 0; c2 < eval::shape::N; ++c2 )
{
M_ret( c1,c2 ) += (*im)( __expr, c1, c2 );
}
}
//perf_mng.stop("3") ;
}
}
perf_mng.stop("cpu") ;
M_cpuTime = perf_mng.getValueInSeconds("cpu");
}
